Digital fraud threatens micro, small, and medium enterprises (MSMEs), yet most automated detection studies remain centered on structured transactions, URLs, emails, or English-language benchmarks. This study evaluates fraud-risk narrative classification in 13,541 deduplicated Indonesian-language texts collected during a three-year window spanning 2023-2026 from Kompas, Detik, YouTube, and X. All retained documents were human-reviewed and assigned a binary discourse-level label and, for risk-positive records, one of seven fraud categories. A tuned TF-IDF linear support vector machine (SVM) was compared with IndoBERT using group-disjoint train-validation-test partitions, validation-only threshold selection, cost-sensitive learning, and 2,000 paired bootstrap resamples. On the untouched test set (n=2,032), IndoBERT achieved 0.9586 Macro-F1, 0.9636 accuracy, 0.9177 Matthews correlation coefficient, and 0.9771 PR-AUC. Its mean Macro-F1 improvement over SVM was 0.0149 with a 95% bootstrap confidence interval of 0.0045-0.0256, whereas the PR-AUC difference was small and uncertain (0.0022; 95% CI: -0.0044-0.0090). IndoBERT reduced false negatives from 69 to 50 and false positives from 31 to 24. Secondary analyses exposed uneven category difficulty, strong topic overlap around social engineering, and a length-sensitive trade-off: IndoBERT was better on short texts but worse than SVM on documents exceeding 512 words. The contribution is an auditable Indonesian MSME-related fraud benchmark that combines duplicate control, a strong classical baseline, calibrated decision thresholds, paired uncertainty, category-level diagnostics, and error analysis rather than proposing a new neural architecture.
Copyrights © 2026